DeepSeek has released dsh-superagent, an MIT-licensed research prototype, presented as a proactive programming agent intended for a single GitHub repository. Packaged for DeepSeek Harness, it listens to GitHub and, optionally, Feishu, then transforms authorized requests into parallel tasks, each in its own agent session.
An agent that waits to be called upon
Its operation is based on the concept of a Ticket, a unit of work opened when an authorized person addresses the agent. Each Ticket runs in its own session and directory, and results are returned to the original conversation, often as a pull request, i.e., a code change proposal submitted for review. A local web interface only lists items requiring human intervention.
The project’s README sets a simple rule: minimize human attention while delivering maximum value. According to the project’s philosophy document, the agent asks for confirmation only once, remains silent during its work, and only speaks again with a result, a question, or a failure.
A Router-Worker architecture
According to the team, common agents consider the session as the basic unit. Citing a report referenced by ifeng, a person can only actively follow 3 to 5 sessions at a time, meaning human attention, not model capability, becomes the limiting factor for parallel work. The proposed architecture has two layers: a long-running Router manages communication and task distribution, while isolated Worker sessions perform the actual work.
The domain model describes a precise sequence: GitHub or Feishu produce events, called Triggers; an evaluation step decides if a batch of Triggers warrants new work; each Ticket carries an immutable Mandate stating why it can be executed; finally, subsequent comments on a pull request are routed to the correct Worker without new mentions. Shell commands move to the background after 5 seconds, so Workers continue to receive messages. Proactive work is governed by three modes: off, observe (default), and auto. In observe mode, the agent only records what it would have accepted.
Results from a simulation
DeepSeek tested the system on simulated group discussion and GitHub CI streams, with about 300 messages per trial, of which about 20 were actual work requests, each configuration run 5 times. The super-agent shows a p50 of 12.4 minutes and a p95 of 15.7 minutes, compared to a p95 of 29.5 minutes for a single session with sub-agents and a neutral prompt, for which no p50 was reported. In this baseline configuration, the sub-agent was not called a single time across 5 trials.
According to the same data, when the arrival rate increased from about 1 to 40 per minute, the batch size increased from 1.0 to 2.8 and the latency per request from 2.5 to 3.9 seconds. These figures are self-reported by DeepSeek and come from secondary accounts of the report. The project also specifies that it was only tested on synthetic messages and simulated tools, and the actual quality of the code produced was not measured.
For installation, the prototype requires Node.js 22.19+ or 24+, git 2.32+, Python 3.8+, tmux, and GitHub CLI 2.48+. The repository finally notes that the agent runs with all the rights of the operating system user, requiring isolation before any use.
